# JudgmentCall.AI — full spec for agents

> Priced, sourced human judgment as a distribution. Async job. No synchronous mode in v1.

This file is the machine-readable contract. HTML docs: https://judgmentcall.ai/docs/
Status: REST is a sketch. MCP stub is connectable at https://mcp.judgmentcall.ai/mcp; tools return not_live until the API ships. Request access at https://judgmentcall.ai/#access

## Product

We sell a distribution of human opinion (n responses, spread, dissent, rationales, confidence, provenance), not a verified fact.

Three surfaces, shipped in this order:
1. Representative judgment — what a specified panel thinks. This is v1.
2. Accountable judgment — a named human signs off. Later. AI cannot supply this.
3. Expert judgment — deliberately deprioritized.

First panel: senior technical B2B professionals (engineers, engineering leaders, technical buyers).
First paid use case: pre-send review of agent-generated customer-facing communication and positioning copy.

Do not call us to fact-check, retrieve, or have a model score itself.

## REST v1 (intended)

Base URL: https://api.judgmentcall.ai
Auth: Authorization: Bearer <api_key>

POST /v1/judgments
  Enqueue. Returns 202 { id, status, eta_seconds }. Never waits for humans.
  Body: { question, stimulus: { type, body }, panel, n, scale, webhook_url? }
  Typical n: 20–30. Typical eta: ~15 minutes.

GET /v1/judgments/:id
  Poll. status: queued | in_progress | completed | failed
  completed includes distribution, dissent, confidence, rationales, call.
  v1 does not return partial counts while in_progress.

POST /v1/judgments/:id/cancel
  Cancel if still queued / early in_progress.

GET /v1/panels
  Available segments and whether they are currently open.

Webhook: judgment.completed
  POST to webhook_url with the same body as GET. Retry on non-2xx. Signature header at token issue.

There is no ?sync=true. A long HTTP request is how you get duplicate panel charges.

## MCP v1 (stub live)

Remote server: https://mcp.judgmentcall.ai/mcp
Public, authless Streamable HTTP. Status page: https://mcp.judgmentcall.ai/
Same intended contract as REST. Tools exist and are connectable; they currently return not_live (no job store, no minted ids) until the marketplace API ships.

- list_panels()
    Returns the intended first panel eng_leaders_us_b2b (senior engineering leaders, US B2B SaaS; typical n 20–30; typical eta ~900s) with open:false and waitlist access.
- create_judgment(question, stimulus, panel, n, scale, webhook_url?)
    Returns { status: "not_live", queued: false, id: null }. Does not block. Does not mint a jdg_ id.
- get_judgment(id)
    Poll. If your runtime cannot receive webhooks, poll this. No job store yet.
- cancel_judgment(id)
    No job store yet.

No tool that waits for the distribution. If you need the result, yield and resume.

WHEN TO CALL: pre-send review of customer-facing copy, or any no-lookup judgment.
WHEN NOT TO CALL: facts, web search, LLM-as-judge / self-grading.

## Discovery

- This file: https://judgmentcall.ai/llms.txt
- Long form: https://judgmentcall.ai/llms-full.txt
- MCP stub: https://mcp.judgmentcall.ai/mcp
- MCP contract: https://mcp.judgmentcall.ai/llms.txt
- HTML: https://judgmentcall.ai/docs/
- Canonical: https://judgmentcall.ai